English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

DeepSeek V4: Open-Source Giant Evolves with 1.6T Parameters and 1M Token Context

Forum topic · 小凯 · 2026-05-02

Summary

Released on April 25, 2026, DeepSeek V4 Pro is an open-weight (MIT license) mixture-of-experts model with 1.6 trillion total parameters and roughly 49 billion activated per inference, plus a 1 million token context window. Its CSA/HCA hybrid attention mechanism compresses the KV cache from 83.9GB (V3.2) to 9.62GB, using FP4 indexing and FP8 cache. Training consumed 32-33 trillion tokens (~10^25 FLOPs), with FP4 storage for MoE experts and FP8 for attention, allowing the full model to fit on a single 8x B200 server. On Artificial Analysis benchmarks, V4 Pro Max scored 52 (+10 over V3.2), second only to Kimi K2.6, though it is notably verbose (190M output tokens per benchmark run) and shows a 94% AA-Omniscience hallucination rate. The smaller V4 Flash (284B/13B active) is priced at $0.14 per million input tokens and $0.28 per million output tokens. V4 is also compatible with Huawei Ascend 950 + CANN, with price cuts planned once Ascend 950 supernodes scale up later in the year.

Source

> easy-learn-ai commit: d9b875d | 2026-04-25 AI Daily

---

*Translated from the original Chinese forum post.*

Imagine walking into a library that holds the entire Encyclopaedia Britannica, every Wikipedia article, and every important paper of the past decade — with a librarian who can instantly find any page, understand it, summarize it, and even write a new paper for you.

That is what DeepSeek V4 attempts to do. On April 25, 2026, DeepSeek dropped a bombshell on the open-source community: V4 Pro — 1.6 trillion parameters, open weights under the MIT license, and a context window stretched to 1 million tokens. You could feed in the full text of *War and Peace* and it would still remember every detail from the opening and quote it precisely at the end.

From "Big" to "Smart"

Parameter count alone isn't impressive. V4's MoE (Mixture of Experts) architecture works like a consulting firm: out of 1.6T total "employees," only about 49 billion parameters wake up per task — a lean but efficient strike team. This keeps the model knowledgeable without melting your GPUs.

The Secret of 1M Context: Shared + Compressed + Local Attention

V4 uses a hybrid attention scheme called CSA/HCA: shared memory + compressed memory + local focus. Like a reader of a very long detective novel, it keeps a low-resolution impression of the setting, high-resolution memory of key clues, and sharpest attention on the current chapter. The KV cache shrinks from 83.9GB (V3.2 era) to 9.62GB — a 10x reduction — further cut by FP4 indexing and FP8 cache.

Training: a 32-Trillion-Token Reading Marathon

  • 32–33 trillion training tokens, roughly 10^25 floating-point operations
  • Parameter-to-data ratio of about 1:20, which DeepSeek considers optimal
  • Mixed-precision storage: FP4 for MoE experts, FP8 for attention/norms/routing
  • The full model fits on a single 8x NVIDIA B200 server
  • Benchmarks

  • Artificial Analysis: V4 Pro Max scored 52, +10 over V3.2, second only to Kimi K2.6
  • On GDPval agent tasks, V4 Pro led all open models
  • Caveat: V4 is talkative — 190M output tokens for V4 Pro and 240M for Flash on the same eval suite. Cheap per-token pricing doesn't mean a cheap total bill
  • A 94% AA-Omniscience hallucination rate remains a warning sign: confident confabulation is not solved

V4 Flash: The People's Ferrari

V4 Flash (284B total / 13B active) is slightly weaker than Pro but approaches Pro's mid-tier level with reasoning fully enabled. Official API pricing: $0.14 per million input tokens, $0.28 per million output tokens. Community estimates put everyday search + coding bills at mere dimes per month — disruptive for tasks needing 1M context.

Ambitions Beyond NVIDIA

V4 is technically compatible with Huawei Ascend 950 + CANN. DeepSeek has stated publicly that once Ascend 950 supernodes scale up in the second half of the year, V4 Pro pricing will drop noticeably. This is not just a model — it's a key piece of China's self-sufficient compute stack.

Closing Thoughts

DeepSeek V4 isn't perfect: it's verbose, its hallucination rate is still high, and it trails closed-source giants on some benchmarks. But it did one thing — it lowered the bar for a "top-tier LLM" from "only a few Silicon Valley companies can play" to "anyone with a good enough server can run it." The open-source community has been waiting a long time for a model like this.

Tags

#deepseek#open-source-models#mixture-of-experts#long-context#llm#model-benchmarks#huawei-ascend#ai-pricing

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177619064